Automated pose estimation for gait analysis in broilers - challenges and insights

Research output: Chapter in Book/Report/Conference proceedingConference paperAcademicpeer-review

Abstract

Computer vision-based approaches are gaining popularity as potential alternatives to traditional gait scoring for phenotyping walking ability of individual broilers on a large scale, facilitating selective breeding. We aim to test whether pose estimation can be used to automatically assess broilers' walking ability. We are developing a pipeline that takes keypoint data as input, and assigns a gait score to individual broilers as output. The pipeline is applied to analyse data from a walking trial, where male broilers were recorded from behind at 14, 21, and 33 days of age as they walked through a corridor (3 m x 0.4 m). We applied a deep learning model, developed in DeepLabCut, to track eight keypoints (head, neck, knees, hocks, feet). We then calculated pose features, keypoint velocities, and accelerations at 12 frames per second. These time series were subjected to machine learning algorithms to classify the walking ability of individual broilers. During this work, we have identified five key challenges in the practical application of pose-estimation based gait analysis in broilers: (1) model transfer, (2) animal identification, (3) quality control of noisy keypoint data, (4) animal visibility and the predictive ability of short sequences, and (5) obtaining a reliable gold standard for gait. We share the lessons learned and propose solutions for these challenges. For example, a pre-trained broiler pose estimation model can be adapted to the new environment with limited further training. The resulting time series will still be noisy and so quality control of the data is a crucial step prior to gait classification using machine learning.

Original languageEnglish
Title of host publication11th European Conference on Precision Livestock Farming
EditorsDaniel Berckmans, Patrizia Tassinari, Daniele Torreggiani
PublisherEuropean Association for Precision Livestock Farming
Pages1350-1357
Number of pages8
ISBN (Electronic)9791221067361
ISBN (Print)9798331303549
Publication statusPublished - Oct 2024
Event11th European Conference on Precision Livestock Farming - Bologna, Italy
Duration: 9 Sept 202412 Sept 2024

Conference/symposium

Conference/symposium11th European Conference on Precision Livestock Farming
Country/TerritoryItaly
CityBologna
Period9/09/2412/09/24

Keywords

  • broiler
  • computer vision
  • gait score
  • lameness
  • machine learning
  • pose estimation

Fingerprint

Dive into the research topics of 'Automated pose estimation for gait analysis in broilers - challenges and insights'. Together they form a unique fingerprint.

Cite this